The spending decision was never the model’s to make.
This post reflects our direct experience operating AI workloads against metered APIs. It is a starting point for discussion — not an accounting recommendation or a procurement policy.
A model built by the company that sells compute will always find a reason that a little more compute is justified, because more compute is what it is for. The reasons it cannot be trusted on the bill stack up, but they all reduce to one picture.
Handing it the spending decision is giving a child a credit card with no budget and asking it to "just buy a few things we need." The child is not malicious. It genuinely believes everything in the cart is needed. It has simply never had to pay the bill, has never been told no, and was raised by the store. The budget exists not because the child is bad, but because the decision was never the child's to make.